cray optimization and performance tools - nersc · • improving performance on hpc systems has...

Cray Optimization and Performance Tools

Harvey Wasserman Woo-Sun Yang

NERSC User Services Group

Cray XE6 Workshop February 7-8, 2011

NERSC Oakland Scientific Facility

Outline

•  Introduction, motivation, some terminology

•  Using CrayPat

•  Using Apprentice2

•  Hands-on lab

2

Why Analyze Performance?

•  Improving performance on HPC systems has compelling economic and scientific rationales. –  Dave Bailey: Value of improving performance of a single application, 5%

of machine’s cycles by 20% over 10 years: $1,500,000 –  Scientific benefit probably much higher

•  Goal: solve problems faster; solve larger problems

•  Accurately state computational need

•  Only that which can be measured can be improved

•  The challenge is mapping the application to an increasingly more complex system architecture –  or set of architectures

3

4

Performance Evaluation as an Iterative Process

Sell Machine

Vendor User

Buy Machine

Improve machine Improve code

Overall goal: more / be.er science results

Performance Analysis Issues

•  Difficult process for real codes •  Many ways of measuring, reporting •  Very broad space: Not just time on one size

–  for fixed size problem (same memory per processor): Strong Scaling

–  scaled up problem (fixed execution time): Weak Scaling

•  A variety of pitfalls abound –  Must compare parallel performance to best

uniprocessor algorithm, not just parallel program on 1 processor (unless it’s best)

–  Be careful relying on any single number •  Amdahl’s Law

5

Performance Questions

•  How can we tell if a program is performing well?

•  Or isn’t?

•  If performance is not “good,” how can we pinpoint why?

•  How can we identify the causes?

•  What can we do about it?

6

7

Supercomputer Architecture

Performance Metrics

•  Primary metric: application time – but gives little indication of efficiency

•  Derived measures: –  rate (Ex.: messages per unit time,

Flops per Second, clocks per instruction), cache utilization

•  Indirect measures: – speedup, efficiency, scalability

8

Performance Metrics

•  Most basic: –  counts: how many MPI_Send calls? –  duration: how much time in MPI_Send ? –  size: what size of message in MPI_Send?

•  (MPI performance as a function of message size)

9

L =Message Size!

T=Time !

}ts = startup cost !

}tw = cost per word!

Tmsg = ts + twL

= Bandwidth!

Performance Data Collection

•  Two dimensions: •  When data collection is triggered:

– Externally (asynchronous): Sampling •  OS interrupts execution at regular intervals and

records the location (program counter) (and / or other event(s))

–  Internally (synchronous): Tracing •  Event based •  Code instrumentation, Automatic or manual

10

Instrumentation

•  Instrumentation: adding measurement probes to the code to observe its execution.

•  Different techniques depending on where the instrumentation is added.

•  Different overheads and levels of accuracy with each technique

11

User-level abstractions problem domain

source code

source code

object code libraries

instrumentation

instrumentation

executable

runtime image

compiler

linker

OS

VM

instrumentation

instrumentation

instrumentation

instrumentation

instrumentation

instrumentation performance data run

preprocessor

Karl Fuerlinger, UCB

Source-Level Instrumentation

•  Goal is to allow performance measurement without modification of user source code

12

Instrumentation

•  Instrumentation: adding measurement probes to the code to observe its execution.

•  Different techniques depending on where the instrumentation is added.

•  Different overheads and levels of accuracy with each technique

13

User-level abstractions problem domain

source code

source code

object code libraries

instrumentation

instrumentation

executable

runtime image

compiler

linker

OS

VM

instrumentation

instrumentation

instrumentation

instrumentation

instrumentation

instrumentation performance data run

preprocessor

Karl Fuerlinger, UCB

Performance Instrumentation

•  Approach: use a tool to “instrument” the code 1.  Transform a binary executable before

executing 2.  Include “hooks” for important events 3.  Run the instrumented executable to capture

those events, write out raw data file 4.  Use some tool(s) to interpret the data

14

Performance Data Collection

•  How performance data are presented: –  Profile: combine sampled events over time

•  Reflects runtime behavior of program entities –  functions, loops, basic blocks –  user-defined “semantic” entities

•  Good for low-overhead performance assessment •  Helps to expose performance hotspots (“bottleneckology”)

–  Trace file: Sequence of events over time •  Gather individual time-stamped events (and arguments) •  Learn when (and where?) events took place on a global timeline •  Common for message passing events (sends/receives) •  Large volume of performance data generated; generally intrusive •  Becomes very difficult at large processor counts, large numbers of

events –  Example in Apprentice section at end of tutorial

15

Performance Analysis Difficulties

•  Tool overhead •  Data overload •  User knows the code better than the tool •  Choice of approaches •  Choice of tools •  CrayPat is an attempt to overcome several

of these – By attempting to include intelligence to identify

problem areas – However, in general the problems remain

16

Performance Tools @ NERSC

•  IPM: Integrated Performance Monitor •  Vendor Tools:

– CrayPat •  Community Tools (Not all fully

supported): – TAU (U. Oregon via ACTS) – OpenSpeedShop (DOE/Krell) – HPCToolKit (Rice U) – PAPI (Performance Application Programming

Interface)

17

Profiling: Inclusive vs. Exclusive

•  Inclusive time for main: – 100 secs

•  Exclusive time for main: – 100-20-50-20=10

secs – Exclusive time

sometimes called “self”

18

USING CRAYPAT

Woo-‐Sun Yang

19

CrayPat Outline

•  Introduction •  Sampling (and example) •  Tracing •  .xf files •  pat_report •  Tracing examples: Heap, MPI, OpenMP •  APA (Automatic Program Analysis) •  CrayPat API •  Monitoring hardware performance counters

•  Exercises provided:

20

Exercise info in this box!

Introduction to CrayPat

•  Suite of tools to provide a wide range of performance-related information

•  Can be used for both sampling and tracing user codes – with or without hardware or network performance

counters

•  Supports Fortran, C, C++, UPC, MPI, Coarray Fortran, OpenMP, Pthreads, SHMEM

21

Access to Cray Tools

•  Access via module utility •  Old:

– module load xt-craypat!– module load apprentice2!

•  Now: – module load perftools – xt-craypat, apprentice2, and xt-papi

(via xt-craypat) are loaded

22

Using CrayPat

1.  Access the tools –  module load perftools!

2.  Build your application; keep .o files –  make clean!–  make!

3.  Instrument application –  pat_build ... a.out!–  Result is a new file, a.out+pat!

4.  Run instrumented application to get top time consuming routines –  aprun ... a.out+pat!–  Result is a new file XXXXX.xf (or a directory containing .xf files)

5.  Run pat_report on that new file; view results –  pat_report XXXXX.xf > my_profile!–  vi my_profile!–  Result is also a new file: XXXXX.ap2

23

Adjust script for +pat

CrayPat Notes

•  Key points to remember:

–  MUST load module prior to building your code •  Error message is obscure! ERROR: Missing required ELF section 'link information'

–  MUST load module prior to looking at man pages

–  MUST run your application in $SCRATCH

–  Module name change: xt-craypat perftools

–  MUST leave relocatable binaries (*.o) when compiling

24

pat_build for Sampling

•  To sample the program counter (PC) at a given time interval or when a specified hardware counter overflows; runs faster than tracing

•  To build, use –S or simply without any tracing flag for pat_build – pat_build –S a.out or – pat_build a.out

25

Running a Sampling Experiment

•  To run – Set PAT_RT_EXPERIMENT to a type

•  Default: samp_pc_time with default time interval (PAT_RT_INTERVAL) of 10,000 microseconds

•  Others: samp_pc_ovfl, samp_cs_time, samp_cs_ovfl (see pat man page)

•  pat_report on .xf from a sampling experiment generates .apa file (later on this)

26

27

Sampling Example > module load perftools!> ftn -c jacobi_serial.f90!> ftn -o jacobi_serial! jacobi_serial.o!> pat_build jacobi_serial!

Run jacobi_serial+pat in a batch job.

> pat_report! jacobi_serial+pat+5511-2558sot.xf!

Exercise in perftools/sampling!

Func@on Group

More on Sampling

•  Binary built for sampling doesn’t work for a tracing experiment

•  Binary built for tracing can be used for sampling (“trace-enhanced sampling”) – Set PAT_RT_EXPERIMENT to a sampling

type – set PAT_RT_SAMPLING_MODE to 1 (raw

user sampling) or 3 (bubble user sampling: aggregate all samples inside a call to a traced function)

28

pat_build for Tracing

•  To trace entries and returns from functions –  Type of experiment: ‘trace’; no need to set

PAT_RT_EXPERIMENT to it in general •  -w: trace functions specified by –t and –T!

–  If none is specified, “trace” the ‘main’ (i.e., entire code as a whole)

•  -u: trace all user functions routine by routine –  For source files owned and writable by the user –  Use care: binary runs longer and can fail

WARNING: Tracing small, frequently called functions can add excessive overhead!

!WARNING: To set a minimum size, say 800 bytes, for traced functions, use:!

!-D trace-text-size=800!

29

pat_build for Tracing

•  -T function: trace function – pat_build –w –T field_,grad_ a.out!– Learn the Unix nm or readelf: nm mycode.o |grep “ T “ !

•  -T !function: do not trace function –  pat_build -u -T \!field_ a.out!

•  trace all user functions except field_ •  ‘\’ to escape the ‘!’ character in csh/tcsh

•  -t tracefile: trace all functions listed in the file tracefile.

30

•  -g: trace all func@ons in certain func@on groups (e.g., MPI): –  pat_build -g mpi,heap -u a.out!

–  pat_build -g mpi -T \!MPI_Barrier a.out trace all MPI calls except MPI_Barrier

•  See $CRAYPAT_ROOT/lib/Trace* for files that list what rou@nes are traced

•  mpi

•  omp •  pthreads •  caf •  upc •  shmem

•  ga •  heap •  blas •  blacs •  lapack

•  scalapack •  fftw •  petsc •  io •  netcdf •  hdf5 •  lustre •  adios •  sysio •  dmapp

•  …

31

CrayPat Trace Function Groups

Other pat_build Options

•  -f: overwrite an existing instrumented program

•  -o instr_prog: –  use a different name for the instrumented

executable instead of a.out+pat –  can put instr_prog at the end of the command

line without ‘-o’ •  -O optfile: use the pat_build options in

the file optfile. – Special argument ‘-O apa’ will be discussed later

Exercise in perftools/pat_build_examples!

32

Instrumenting Programs Using Compiler Options

•  Available in Pathscale, GNU, and Cray compilers at NERSC

•  Requires recompile, link (selected files) •  Alternative to pat_build –u

–  GNU, PathScale: cc –finstrument-functions -c pgm.c! cc –o pgm pgm.c!

–  Cray compiler: cc –h func_trace -c pgm.c! cc –o pgm pgm.o

–  Then pat_build –w pgm!

33

.xf Files

•  Experiment data files; binary files •  Number of .xf files

–  Single file (for ≤ 256 PEs) or directory containing multiple (~√PEs) files

–  Can be changed with PAT_RT_EXPFILE_MAX •  Name convention

–  a.out+pat+<UNIX_PID>-<NODE_ID>[st][dfot].xf (or a.out+apa+….xf; see APA)

–  [st][dfot]: [sampling, tracing], [distributed memory, forked process, OpenMP, Pthreads]

•  New one each time you run your application •  Can create .xf file(s) in other than the current

location by setting PAT_RT_EXPFILE_DIR

34

pat_report

•  Generates from .xf data file(s) ASCII text report and .ap2 file (to be viewed with Apprentice2) –  Create .ap2 file right after .xf file becomes available!

•  .xf file requires the instrumented executable in the original directory (not portable)

•  .ap2 doesn’t (self-contained and portable) •  pat_report on .xf file (or directory containing

multiple .xf files) generates –  text report to stdout (terminal) –  .ap2 file –  .apa file, in case of a sampling experiment

•  Running on .ap2 file generates text report to stdout

35

pat_report Options

•  -d: data items to display (time data, heap data, counter data,…)

•  -b: how data is aggregated or labeled (group, function, pes, thread, …)

•  -s: details of report appearance (aggregation, format,…)

•  -O|-b|-d|-s –h: list all available cases for the option!

36

pat_report Options

•  -O: predefined report types; this is what we should use!–  profile, callers (ca), calltree (ct), ca+src, ct+src, heap, load_balance, mpi_callers, hwpc, nids, …

–  heap and load_balance have a few “sub-reports” •  load_balance = load_balance_program (for entire program)

+ load_balance_group (per function group) + load_balance_function (per function)

–  Examples: -O profile: -d ti%@0.95,ti,imb_ti,imb_ti%,tr! -b gr,fu,pe=HIDE!-O callers+src: -d ti%@0.95,ti,tr -b gr,fu,ca,pe=HIDE! –s show_ca=‘fu,so,li’!-O load_balance_function: -d ti%@0.95,ti,tr!! ! ! ! ! ! ! -b gr,fu,pe=[mmm]!

37

pat_report Options

•  Without –d, -b or -O, a few reports appear by default; dependent on the used trace groups

•  -i instr_prog: specify the path for the instrumented executable (if not in the same directory as the .xf file)

•  -o output_file: specify the output file name •  -T: disable all thresholds (5%)

•  pat_report lists the options used in the report – a good place to learn options; try adding an option to the existing ones

•  By default, all reports (-O) show either no individual PE values or only the PEs having the maximum, median, and minimum values.

•  The suffix _all can be appended to any of the pat_report keyword options to show the data for all PEs

38

Exercise in perftools/heap!

39

Heap Memory Example > module load perftools!> ftn -c jacobi_serial.f90!> ftn -o jacobi_serial ! jacobi_serial.o!> pat_build -g heap -u! jacobi_serial!

Run jacobi_serial+pat in a batch job.

> pat_report! jacobi_serial+pat+15243-18tot.xf!

Some@mes not easy to understand

Heap func@on group

•  Profiling by MPI func@ons

•  MPI message stats

•  load imbalance among MPI tasks

> module load perftools!> ftn -c jacobi_mpi.f90!> ftn -o jacobi_mpi! jacobi_mpi.o!> pat_build -g mpi -u jacobi_mpi!

Run jacobi_mpi+pat in a batch job.

> pat_report! jacobi_mpi+pat+15207-18tdt.xf!

40

MPI Code Example

Exercise in perftools/mpi!

per PE

41

MPI Code Example imb = max – avg imb% = imb/max * npes/(npes-1) * 100%

No per-‐PE info PEs with max, min, median values

Func@on Groups

42

MPI Code Example

Bins by message size

per PE

Level of depth

callers

•  Time spent wai@ng at a barrier before entering a collec@ve can be a significant indica@on of load imbalance.

•  MPI_SYNC group: for @me spent wai@ng at the barrier before entering the collec@ves

•  Actual @me spent in the collec@ves go to the MPI func@on group

•  Not to separate these groups, set PAT_RT_MPI_SYNC to 0 before aprun

43

MPI_SYNC Function Group

•  ‘-O mpi_sm_rank_order!![-s rank_grid_dim=M,N]!![-s rank_cell_dim=m,n]!![-s mpi_dest=d]’:

–  Based on sent messages –  -s rank_*: specify a

different MPI process topology »  Global topology, M×N!»  topology per node, m×n!

–  consider d busiest partners (default, 8)

MPI Rank Order Suggestion

44

> pat_report -O mpi_sm_rank_order jacobi_mpi+pat+30971-19tdot.xf > ls MPICH_RANK_ORDER.* MPICH_RANK_ORDER.d MPICH_RANK_ORDER.u

Examined the cases 0, 1, 2, and 3 (‘d’ and ‘u’) for MPICH_RANK_REORDER_METHOD; provided MPICH_RANK_ORDER file for the ‘d’ and ‘u’ cases.

Best: smallest ra@os wrt to SMP ord.


45

•  ‘-O mpi_rank_order !![-s mro_metric=…]’: based on specified metric

–  ‘time’, if no metric is specified –  HWPC if a.out+pat was run

with this set (later on this)!


46

> pat_report -O mpi_rank_order jacobi_mpi+pat+30971-19tdot.ap2 > ls MPICH_RANK_ORDER.* MPICH_RANK_ORDER.d

Examined the cases 0, 1, 2, and 3 (‘d’) for MPICH_RANK_REORDER_METHOD; provided MPICH_RANK_ORDER file for ‘d’ case.

> module load perftools!> ftn -mp=nonuma -c jacobi_omp.f90!> ftn -mp=nonuma -o jacobi_omp! jacobi_omp.o!> pat_build -g omp -u jacobi_omp!

Run jacobi_omp+pat in a batch job

> pat_report! jacobi_omp+pat+15307-18tot.xf!

Exercise in perftools/openmp!

OpenMP Code Example

47

48

OpenMP Code Example

•  24 thread case •  Good load balance

in parallel regions!

•  Overhead: –  Imb. =100%!, all in

the master thread (pat_report’s ‘–O load_balance’ to see this more clearly)

–  largest in init_fields_

No per-‐thread info

49

OpenMP Code Example

jacobi_omp.f90 with ngrind=9,600 and maxiter=20

Automatic Program Analysis (APA)

•  One may not know in advance where large run time is spent; tracing all the functions can be overwhelming due to large overhead

1.  Have the tool detect the most time consuming functions in the application with a sampling experiment

2.  Feed this information back to the tool to instrument for focused data collection

3.  Get performance information on the most significant parts of the application

•  APA does this for you (you can do the same thing by hand)

50

Automatic Program Analysis (APA)

1.  pat_build -O apa a.out •  Produces the instrumented executable a.out+pat for sampling

2.  aprun -n … a.out+pat •  Produces data file, e.g., a.out+pat+4677-19sdot.xf

3.  pat_report a.out+pat+4571-19sdot.xf •  Produces a.out+pat+4571-19sdot.apa (suggested options for tracing) •  Produces a.out+pat+4571-19sdot.ap2

4.  Edit a.out+pat+4571-19sdot.apa, if necessary (next slide) 5.  pat_build -O a.out+pat+4571-19sdot.apa

•  Produces a.out+apa for tracing 6.  aprun -n … a.out+apa

•  Produces a.out+apa+4590-19tdot.xf 7.  pat_report a.out+apa+4590-19tdot.xf > out

51

Exercise in perftools/apa!

•  Recommended pat_build op@ons for tracing

•  Customize it for your need –  Include/exclude funcTons –  Add/change opTons

•  Note that the sugges@ons may not be valid for a very different task/thread configura@on

52

.apa File

PAT_RT_HWPC set to 1

mpi trace group chosen

Trace ‘MAIN_’, but not ‘compute_diff_’

Instrumented program to create Program to instrument

Command to run

Sampling run result

CrayPat Application Program Interface (API)

•  Assume your code contains initialization and solution sections

•  Want to analyze performance only of solution

•  How to do this? Several approaches: –  Init section is only one routine (or just a few

routines): eliminate it (or them) from the profile. –  Init section is many routines: Use API to define a

profile region that excludes init –  What happens if some routines shared by init

and solve? Use API to turn profiling on and off as needed

53

Exercise in perftools/api!

Arbitrary, user-assigned id’s

54

Using the API > module load perftools!> ftn -c jacobi_mpi_api.f90!> ftn -o jacobi_mpi_api jacobi_mpi_api.o!> pat_build -g mpi -u jacobi_mpi_api!

Run jacobi_mpi_api+pat in a batch job. !> pat_report! jacobi_mpi_api+pat+502-19tdot.xf!

Hardware Performance Counters

•  Registers available on the processor that count certain events

•  Minimal overhead – They’re running all the time – Typically one clock period to read

•  Potentially rich source of performance information

55

Types of Counters

•  Cycles •  Instruction count •  Memory references, cache hits/misses •  Floating-point instructions •  Resource utilization

56

PAPI Event Counters

•  PAPI (Performance API) provides a standard interface for use of the performance counters in major microprocessors

•  Predefined actual and derived counters supported on the system –  To see the list, run ‘papi_avail’ on compute node via aprun:

module load perftools!!!aprun –n 1 papi_avail!

•  AMD native events also provided; use ‘papi_native_avail’:

!! !aprun –n 1 papi_native_avail

57

Hardware Performance Monitoring

•  Specify hardware counters to be monitored during sampling or tracing –  Default is “off” (no HW counters measured) –  Choose up to 4 events

•  Can specify individual events: setenv PAT_RT_HWPC “PAPI_FP_OPS,PAPI_L1_DCM”! aprun –n … a.out+pat (or a.out+apa)

•  Or predefined event group number (next slide): setenv PAT_RT_HWPC 1! aprun –n … a.out+pat (or a.out+apa)

•  Multiplexing (monitoring more than 4 events) to be supported in later versions (5.2?)

58

Exercise in perftools/hwpc!

Predefined Counter Groups for PAT_RT_HWPC

0 Summary with instruction metrics 1 Summary with translation lookaside buffer (TLB) metrics 2 L1 and L2 cache metrics 3 Bandwith information 4 *** DO NOT USE, not supported on Quad-core or later AMD Opteron processors *** 5 Floating point instructions 6 Cycles stalled and resources empty 7 Cycles stalled and resources full 8 Instructions and branches 9 Instruction cache values 10 Cache hierarchy 11 Floating point instructions (2) 12 Floating point instructions (vectorization) 13 Floating point instructions (single precision) 14 Floating point instructions (double precision) 15 L3 cache 16 L3 cache, core-level reads 17 L3 cache, core-level misses 18 L3 cache, core-level fills caused by L2 evictions 19 Prefetches

59


60

PAT_RT_HWPC = 1

Measured

Derived

avg uses (or hits): per word per miss

pat_report –s data_size=4 … (because single precision was used)

cacheline: 64 Bytes L1 cache: 64 KB, dedicated for each core L2 cache: 512 KB, dedicated for each core page: 4 KB (2 MB if huge page)


61

PAT_RT_HWPC = 2


62

PAT_RT_HWPC = 5

RelaTve raTos for mulTplies and adds


63

PAT_RT_HWPC = 12

vector length (for sp) for 128-‐bit wide SSE2 vector operaTon = 16583040480 / 4154398320 = 3.99 Compiled with ‘`n –fastsse …’

Add and mulTply instrucTons issued

Adds and mulTplies performed

Guidelines to Identify the Need for Optimization

64

* Suggested by Cray

Derived metric Op@miza@on needed when* PAT_RT_HWPC

ComputaTonal intensity < 0.5 ops/ref 0, 1

L1 cache hit raTo < 90% 0, 1, 2

L1 cache uTlizaTon (misses) < 1 avg hit 0, 1, 2

L1+L2 cache hit raTo < 92% 2

L1+L2 cache uTlizaTon (misses) < 1 avg hit 2

TLB uTlizaTon < 0.9 avg use 1

(FP MulTply / FP Ops) or (FP Add / FP Ops)

< 25% 5

VectorizaTon < 1.5 for dp; 3 for sp 12 (13, 14)

Monitoring Network Performance Counters

•  Use PAT_RT_NWPC instead of PAT_RT_HWPC

•  See ‘Overview of Gemini Hardware Counters’, S-0025-10 – http://docs.cray.com

65

USING CRAY’S APPRENTICE TOOL

Harvey Wasserman

66

Using Apprentice

•  Optional visualization tool for Cray’s perftools data

•  Use it in a X Windows environment •  Uses a data file as input (XXX.ap2) that is

prepared by pat_report!1.   module load perftools!2.   ftn -c mpptest.f!3.   ftn -o mpptest mpptest.o!4.   pat_build -u -g mpi mpptest!5.   aprun -n 16 mpptest+pat!6.   pat_report mpptest+pat+PID.xf >

my_report!7.   app2 [--limit_per_pe tags] [XXX.ap2]!

67

Opening Files

•  Identify files on the command line or via the GUI:

68

69

Apprentice Basic View Can select new

(addiTonal) data file and do a screen dump

Can select other views of the data

Worthless Useful

Can drag the “calipers” to focus the view on porTons of the run

70

Apprentice Call Tree Report Horizontal size =

cumulaTve Tme in node’s children

VerTcal size = Tme in computaTon

Green nodes: no callees

Stacked bar charts: load balancing info. Yellow=Max

purple=Average Light Blue=Minimum

Calipers work

Right-‐click to view source

Useful

71

Apprentice Call Tree Report Red arc idenTfies path to the highest detected load

imbalance.

Call tree stops there because nodes were filtered out. To see the

hidden nodes, right-‐click on the node a.ached to the marker and select "unhide all children” or "unhide one

level".

Double-‐click on for more info about load

imbalance.

Apprentice Event Trace Views

•  Run code with setenv PAT_RT_SUMMARY 0

•  Caution: Can generate enormous data files and take forever

72

Apprentice Traffic Report

73

Shows message traces as a funcTon of Tme

Look for large blocks of barriers held up by a single

processor

Zoom is important; also, run just a porTon of your

simulaTon

Scroll, zoom, filter: right-‐click on trace

Click here to select this report

Apprentice Traffic Report: Zoomed

•  Mouse hover pops up window showing source location. 74

75

Tracing Analysis Example

Mosaic View

• 

• 

76

Click here to select this report

Can right-‐click here for more opTons

Colors show average Tme (green=low, red=high)

Very difficult to interpret by itself – use the Craypat

message staTsTcs with it.

Shows Interprocessor communicaTon topology and color-‐coded intensity

77

Mosaic View

SP CG

LU

MG

FT BT

NERSC6 Application Benchmark Characteristics

Benchmark Science Area Algorithm Space Base Case Concurrency

Problem Description

CAM Climate (BER) Navier Stokes CFD 56, 240 Strong scaling

D Grid, (~.5 deg resolution); 240 timesteps

GAMESS Quantum Chem (BES)

Dense linear algebra 384, 1024 (Same as Ti-09)

DFT gradient, MP2 gradient

GTC Fusion (FES) PIC, finite difference 512, 2048 Weak scaling

100 particles per cell

IMPACT-T Accelerator Physics (HEP)

PIC, FFT component 256,1024 Strong scaling

50 particles per cell

MAESTRO Astrophysics (HEP)

Low Mach Hydro; block structured-grid multiphysics

512, 2048 Weak scaling

16 32^3 boxes per proc; 10 timesteps

MILC Lattice Gauge Physics (NP)

Conjugate gradient, sparse matrix; FFT

256, 1024, 8192 Weak scaling

8x8x8x9 Local Grid, ~70,000 iters

PARATEC Material Science (BES)

DFT; FFT, BLAS3 256, 1024 Strong scaling

686 Atoms, 1372 bands, 20 iters

78

NERSC6 Benchmarks Communication Topology*

MILC

PARATEC IMPACT-‐T CAM

MAESTRO GTC

79 *From IPM

Sample of CI & %MPI

*CI is the computaTonal intensity, the raTo of # of FloaTng Point OperaTons to # of memory operaTons.

80

For More Information

•  Using Cray Performance Analysis Tools, S–2376–51 –  http://docs.cray.com/books/S-2376-51/S-2376-51.pdf

•  man craypat •  man pat_build •  man pat_report •  man pat_help very useful tutorial program •  man app2 •  man hwpc •  man intro_perftools •  man papi •  man papi_counters

81

For More Information

•  “Performance Tuning of Scientific Applications,” CRC Press 2010

82

83

Exercise

Same code, same problem size, run on the same 24 cores. What is different? Why might one perform be.er than the other? What performance characterisTcs are different?

Exercise

•  Get the sweep3d code. Untar •  To build: type ‘make mpi’ •  Instrument for mpi, user •  Get an interactive batch session, 24 cores •  Run 3 sweep3d cases on 24 cores creating

Apprentice traffic/mosaic views: –  cp input1 input; aprun –n 24 …!–  cp input2 input; aprun –n 24 …!–  cp input3 input; aprun –n 24 …!

•  View the results from each run in Apprentice and try to explain what you see.

84

85

Performance Metrics

CPU Time = Ninst * CPI * Clock rate

Application

Compiler

CPU Time = Instructions --------------- Program

Cycles ------------- Instruction

Seconds ------------- Cycle

X X

Instruction Set

Architecture

Technology

cray optimization and performance tools - nersc · • improving performance on hpc systems has...

Documents